Vertex AI & pgvector search enabled

Video
Intelligence Platform

Dynamically sync video archives, generate rich semantic descriptions from transcripts, and privately index listings. Engineered with Next.js 16, Prisma 7, and Gemini 3.5 Flash.

Explore Ingestion Technical Specs
1 Quota
Playlist Sync Cost
3.5 Flash
Auto-Tagging Model
RRF Hybrid
Search Precision
ingestion-pipeline.sh
ACTIVE
$python3 src/main.py --sync-uploads
[INFO] Starting uploads playlist synchronization...
[INFO] YouTube items discovered (cost: 1 quota unit).
[SUCCESS] Sync complete. Stored 150 video shell definitions.
$python3 src/main.py --enrich --video-id "dQw4w9WgXcQ"
>> Calling Gemini 3.5 Flash for automated layout & tag extraction...
[AI SUCCESS] Structured output parsed successfully:
Indexing: Walkthrough Unit #14C (2BR, Laundry, Balcony)
[DB] Upsert completed. Local pgvector database is 100% synchronized.
Platform Interface

Premium Matte Black Studio

Experience a seamless 1:1 YouTube-inspired desktop experience with live dynamic metrics and natural unit ordering.

ListingVault Studio Dashboard
Real Estate Recall

Decaying Metadata vs.
Semantic Intelligence

Finding the correct client walkthrough from thousands of generic uploads without revealing unit addresses publicly.

The Friction

Decaying Video Archives

Uploading generic titles to YouTube publicly to preserve privacy makes indexing impossible over time, leading to severe recall friction.

  • Generic titles ("Walkthrough #1") obscure the real address.
  • YouTube API default search consumes an expensive 100 quota units.
  • Lost context: older property uploads are hard to retrieve.
The Innovation

Private Video Cataloging

Synchronize channels cheaply, extract semantic details silently via Gemini, and search privately inside a premium localized Studio.

  • Private unit details, layout, and tags stored securely.
  • upload playlist indexing costs exactly 1 API quota unit.
  • Semantic hybrid search combined with Natural Unit sorting.
System Capabilities

High Performance
Architecture Bento

Uploads Playlist API Strategy

Bypasses the expensive 100 quota unit searches. Dynamically crawls YouTube's `playlistItems` recursively, costing exactly 1 quota unit to backfill thousands of walkthrough records instantly.

Gemini 3.5 Flash Analysis

Calls Vertex AI global endpoint with structured JSON Schema outputs to tag and catalog property configurations, layouts (Studio, 1BR), and amenities seamlessly.

Gemini Embeddings v2 (GA)

Leverages the cutting-edge gemini-embedding-2 model, Google's unified multimodal champion. Mapping text-only metadata into optimized 768-dim vector segments yields 30% higher semantic recall precision, backed by localized offline all-MiniLM-L6-v2 engines for instant offline development parity.

Natural Unit Sorting

Advanced string parsing on layout unit lists, placing Unit 6 naturally before Unit 10, resolving standard ASCII lexicographical sorting errors.

Secure NextAuth Middleware

Guards index ingestion and search queries using standard NextAuth session providers, locking private real-estate database catalogs from public viewing.

Pipeline Flow

E2E Automated Ingestion

1

OAuth 2.0 Sync

Authorize with Google via NextAuth to securely fetch video feeds viauploads playlist.

2

AI Content Analysis

Gemini 3.5 Flash parses description blocks to extract unit layouts and tags privately.

3

Embedding Generation

Generate unified multimodal semantic vectors using Google's next-generation gemini-embedding-2 model.

4

Hybrid Search Index

Combine tsvector keywords and pgvector semantic RRF scores to present walkthroughs instantly.